COMPUTATIONAL EFFICIENCY OF AN ARTIFICIAL NEURAL NETWORK IN OPTIMIZING THE OPERATING CONDITIONS OF PATCH ANTENNAS
Radio physics and radio astronomy · 2026 · 31 (2) · 098-107
DOI: 10.15407/rpra31.02.098 ↗Kurz gefasst
Fachartikel in der Zeitschrift Radio physics and radio astronomy, 2026. Verfasst von S. Butov und 5 weiteren Autor:innen. Die Arbeit nennt 27 Literaturangaben.
Abstract (Originalsprache)
Subject and Purpose. Designing antennas with desired operational features is a complex optimization problem addressing a large number of parameters and nonlinear relationships. Artificial neural networks (ANNs) have had strong potential in antenna engineering. The ability to approximate nonlinear functions and capture hidden relationships enables partial automation of antenna designing. However, it requires a detailed analysis of how ANN parameters affect predictive accuracy and computational efficiency. The present study investigates the influence of ANN architecture and training configurations on the predictive accuracy of resonant frequencies for rectangular microstrip patch antennas. Methods and Methodology. A modular Python-based experimental platform is used to evaluate ANN performance for dif- ferent ANN configurations. The ANN training is on a synthetic dataset generated from a classical analytical antenna model. The number of hidden layers, neurons per layer, activation functions, and optimizers are systematically varied to assess their particular impacts on convergence, generalization performance, and execution time.
Results
It has been shown that a three-layer neural network [1024, 512, 256 neurons] with ReLU activation and the Adam optimizer strikes the best balance between predictive accuracy and training rate. Simpler or excessively deep architectures, non-adaptive optimizers, and saturating activation functions can slow convergence or cause unstable training. Further analysis indicated that a smaller batch size introduces useful stochasticity into training but might also destabilize the process.
Conclusions
The study has demonstrated that joint optimization of ANN architecture and training dynamics is essential for developing accurate and computationally efficient electromagnetic models. The practical results are recommendations for machine-learning-based antenna design.
Zitieren
APA
Butov, S., Tuz, A., Morozova, O., Vinnichenko, S., Khardikov, O., & Shulga, S. (2026). COMPUTATIONAL EFFICIENCY OF AN ARTIFICIAL NEURAL NETWORK IN OPTIMIZING THE OPERATING CONDITIONS OF PATCH ANTENNAS. Radio physics and radio astronomy, 31(2), 098-107. https://doi.org/10.15407/rpra31.02.098
Vancouver
Butov S, Tuz A, Morozova O, Vinnichenko S, Khardikov O, Shulga S. COMPUTATIONAL EFFICIENCY OF AN ARTIFICIAL NEURAL NETWORK IN OPTIMIZING THE OPERATING CONDITIONS OF PATCH ANTENNAS. Radio physics and radio astronomy. 2026;31(2):098-107. doi:10.15407/rpra31.02.098
BibTeX
@article{Butov2026,
author = {Butov, S. and Tuz, A. and Morozova, O. and Vinnichenko, S. and Khardikov, O. and Shulga, S.},
title = {COMPUTATIONAL EFFICIENCY OF AN ARTIFICIAL NEURAL NETWORK IN OPTIMIZING THE OPERATING CONDITIONS OF PATCH ANTENNAS},
journal = {Radio physics and radio astronomy},
year = {2026},
volume = {31},
number = {2},
pages = {098--107},
doi = {10.15407/rpra31.02.098}
}
EUVIMED